Using Some Metric Distance in Local Density Based on Outlier Detection Methods

Abstract

This study used different metric distances to estimate density functions in outlier detection. We employed multidimensional scaling for dimension reduction using two metric distances (the standardized Euclidean and Minkowski distances).  A local density-based method was applied to three methods for outlier detection. We use the criterion for evaluating the performance of outlier approaches in this paper is Precision. The Gaussian local density estimation method uses three nearest neighbors types (KNN, RNN, and SNN). While in SGR, and Volcano use one kind of nearest neighbor (KNN). Extensive experiments on a synthetic dataset have shown that the result of the two distances was approximately equal. The RDOS and the VOL methods are more efficient when we increase the number of nearest neighbours. The average numbers of outliers increase in the SGR method, when we grow NN, the average number of outliers appears weak in the technique.